Design and Implementation of a High-Level AI Architecture for Stock price prediction

Abstract

This research provides a comprehensive evaluation of machine learning effectiveness in financial markets, evolving from basic regression to a complex multi-modal architecture. The analysis effectively exposes the "data shuffling trap," demonstrating how traditional evaluation metrics can lead to severe look-ahead bias if chronological order is not strictly preserved. By incorporating FinBERT for sentiment analysis, the work successfully addresses the inherent limitations of historical price data, which often fails to account for external catalysts. A standout contribution is the Hybrid Momentum Index (HMI), a framework that synthesizes technical trends, volatility, and qualitative sentiment into a unified predictive engine. The NVDA case study serves as a strong empirical validation, identifying contrarian signals that would be invisible to standard algorithmic models. Ultimately, this thesis offers a robust synthesis of market "physics" and investor psychology, significantly advancing the discussion on AI-driven financial forecasting.